TDR: Two-stage deep recommendation model based on mSDA and DNN

Recently, deep learning techniques have been widely used in recommendation tasks and have attained record performance. However, the input quality of the deep learning model has a great influence on the recommendation performance. In this work, an efficient and effective input optimization method is proposed. Specifically, we propose an integrated recommendation framework based on two-stage deep learning. In the first stage, with user and item features as the original input, a low-cost marginalized stacked denoising auto-encoder (mSDA) model is used to learn the latent factors of users and items. In the second stage, the resulting latent factors are combined and used as input vector to the DNN model for fast and accurate prediction. Using the latent factor vector as the input to the deep learning-based recommendation model not only captures the high-order feature interaction, but also reduces the burden of the hidden layer, and also avoids the model training falling into local optimum. Extensive experiments with real-world datasets show that the proposed model shows much better performance than the state-of-the-art recommendation methods in terms of prediction accuracy, parameter space and training speed.